Deep Boosting Robustness of DNN-based Image Watermarking via DBM ARK Guanhui YeJiashi GaoWei XieBo Yin Xuetao Wei

2025-05-06 0 0 6.92MB 9 页 10玖币

侵权投诉

Deep Boosting Robustness of DNN-based Image Watermarking via DBMARK

Guanhui Ye∗Jiashi Gao∗Wei Xie◦Bo Yin⊗Xuetao Wei∗

∗Southern University of Science and Technology ◦Hunan University ⊗Changsha University of Science and Technology

Abstract

Image watermarking is a technique for hiding informa-

tion into images that can withstand distortions while re-

quiring the encoded image to be perceptually identical to

the original image. Recent work based on deep neural net-

works (DNN) has achieved impressive progression in digi-

tal watermarking. Higher robustness under various distor-

tions is the eternal pursuit of digital image watermarking

approaches. In this paper, we propose DBMARK, a novel

end-to-end digital image watermarking framework to deep

boost the robustness of DNN-based image watermarking.

The key novelty is the synergy of invertible neural networks

(INN) and effective watermark features generation. The

framework generates watermark features with redundancy

and error correction ability through the effective neural net-

work based message processor, synergized with the power-

ful information embedding and extraction abilities of INN

to achieve higher robustness and invisibility. The power-

ful learning ability of neural networks enables the message

processor to adapt to various distortions. In addition, we

propose to embed the watermark information in the discrete

wavelet transform (DWT) domain and design low-low (LL)

sub-band loss to enhance invisibility. Extensive experiment

results demonstrate the superiority of the proposed frame-

work compared with the state-of-the-art ones under various

distortions such as dropout, cropout, crop, Gaussian ﬁlter,

and JPEG compression.

1. Introduction

Digital watermarking has been widely used in the copy-

right protection of multimedia products since its inception

[29]. Digital watermarking hides the watermark informa-

tion with speciﬁc meanings in digital content, such as im-

ages, videos, audio, documents, etc., through digital em-

bedding. The extraction and recovery of watermark infor-

mation can be used to prove the ownership and as evidence

for identifying illegal infringement. The goal of digital wa-

termarking is to embed the secret message into the cover

image in an invisible way and to extract the accurate secret

message in the case of various distortions. In other words,

digital watermarking requires high robustness and high in-

visibility. Least signiﬁcant bits (LSB) [29] was the earliest

research on image information hiding, which encodes the

secret message on the least signiﬁcant bits of image pix-

els. However, statistical measures [10–12] can easily detect

the secret information hidden by LSB. Furthermore, the re-

searchers ﬁnd that watermarking in the frequency domain

is more robust than the spatial ones. However, these tra-

ditional methods are heavily dependent on shallow manual

image features, which imply that they need to be carefully

designed and can not fully use the redundant information of

cover images, so the robustness of such methods is limited.

In recent years, with the upsurge of deep learning, many

researchers have applied deep neural networks (DNN) to

digital image watermarking, which signiﬁcantly facilitates

its development. These DNN-based methods [17,22,35]

have shown advantages in robustness under various dis-

tortions compared with traditional methods. Zhu et al.

[35] proposed the ﬁrst DNN-based method named Hid-

den and demonstrated superior performance than most tra-

ditional methods. Meanwhile, various subsequent DNN-

based methods have adopted a similar framework. Such

a framework uses a separate encoder and decoder, which

treats the watermark encoding and decoding processes in-

dependently. Xu et al. [32] simply applied invertible neu-

ral networks (INN) in image watermarking, which did not

achieve satisfactory performance. Since higher robustness

is the eternal pursuit of these DNN-based image watermark-

ing approaches, our research question is: how to deep boost

the robustness of DNN-based image watermarking under

various distortions?

In this paper, we propose a novel end-to-end digital im-

age watermarking framework DBMARK to deep boost the

robustness of DNN-based image watermarking. The key

novelty is the synergy of invertible neural networks (INN)

and effective watermark features generation. The frame-

work generates watermark features with redundancy and

error correction ability through the effective neural net-

work based message processor, synergized with the power-

ful information embedding and extraction abilities of INN

to achieve higher robustness and invisibility. The power-

ful learning ability of neural networks enables the message

arXiv:2210.13801v3 [cs.CV] 16 Nov 2022

Figure 1. The architecture of DBMARK. In the watermark encoding process, a secret message Mis encoded into a cover image Ico to

generate an encoded image Ien. In watermark decoding process, the encoded image Ien passing through the noise layer is fed in a reverse

direction to recover the secret message M0. Here, we use dense block in our φ(·),ρ(·)and η(·)modules. The forward process and the

reverse process of INN share the same network parameters and modules.

processor to adapt to various distortions. We embed and

extract watermark features by INN, in which the forward

and inverse processes of INN share the same parameters

and modules. Instead of hiding information directly in the

spatial domain, we embed watermark information in the fre-

quency domain of the cover image through the DWT. More-

over, we embed most of the watermark information into the

high-frequency component through the LL sub-band loss to

improve invisibility from the perspective of the human vi-

sual system. Experiment results show that our DBMARK

framework outperforms existing SOTA (State Of The Art)

methods on robustness and invisibility evidently.

In summary, the main contributions of this paper are:

• We present a novel end-to-end digital image water-

marking framework DBMARK to deep boost the ro-

bustness of DNN-based image watermarking, which

is the synergy of the invertible neural networks (INN)

and effective watermark features generation.

• We propose the neural network based message proces-

sor to generate watermark features with redundancy

and error correction ability, signiﬁcantly improving ro-

bustness against various distortions simultaneously.

• We propose low-low (LL) sub-band loss to embed

more watermark information in the high-frequency

component of the discrete wavelet transform (DWT)

domain, which enhances the invisibility of our DB-

MARK evidently.

• We conduct extensive experiments to demonstrate that

our framework DBMARK achieves higher robustness

and invisibility than state-of-the-art approaches under

various distortions.

The rest of the paper is organized as follows. We re-

view the related work about digital watermarking and INN

in Section 2. The details of the proposed framework are de-

scribed in Section 3. Extensive experiments are presented

in Section 4. Finally, we conclude our work in Section 5.

2. Related Work

2.1. Digital Watermarking

As a primary technology for copyright protection of con-

tent, digital watermarking is a popular research area in a

wide range of real-world scenarios [5,7,14,19,20,26]. The

prior research of traditional methods has mainly investi-

gated pixel-level manipulation [4,29] in spatial domain. In

order to improve the robustness, the traditional methods tar-

geted the frequency domain, such as discrete Fourier trans-

form (DFT) [27] domain, discrete cosine transform (DCT)

[15] domain and discrete wavelet transform (DWT) [13] do-

main.

In recent years, DNN-based methods have shown more

advantages than traditional methods in both invisibility and

robustness against various distortions, which is due to the

powerful feature extraction ability of deep neural networks.

Zhu et al. [35] ﬁrst proposed a DNN-based framework to

jointly trained the encoder and the decoder with a noise

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

DeepBoostingRobustnessofDNN-basedImageWatermarkingviaDBMARKGuanhuiYeJiashiGaoWeiXieBoYinXuetaoWeiSouthernUniversityofScienceandTechnologyHunanUniversityChangshaUniversityofScienceandTechnologyAbstractImagewatermarkingisatechniqueforhidinginforma-tionintoimagesthatcanwithstanddistortionswhilere...

展开>> 收起<<

Deep Boosting Robustness of DNN-based Image Watermarking via DBM ARK Guanhui YeJiashi GaoWei XieBo Yin Xuetao Wei.pdf

共9页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Deep Boosting Robustness of DNN-based Image Watermarking via DBM ARK Guanhui YeJiashi GaoWei XieBo Yin Xuetao Wei

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: